JP4318475B2 - Speaker authentication device and speaker authentication program - Google Patents

Description

[0001]
BACKGROUND OF THE INVENTION
The present invention relates to a speaker authentication device and a speaker authentication program for specifying who is a speaker who emits voice.
[0002]
[Prior art]
A speaker authentication device that authenticates who is a speaker from a keyword uttered by a user and performs security management based on the authentication result is used.
[0003]
In such a speaker authentication device, it is necessary to register the feature quantity of the voice signal of a keyword (such as a personal identification number) specifying a user in association with the user prior to the authentication as a verification database. The feature amount of the audio signal corresponding to this keyword is used as verification information when performing user authentication. When the user is to be authenticated, the user inputs voice including a keyword registered in advance as authentication information to the speaker authentication apparatus. By comparing the authentication information with the verification information, the verification information having a high similarity with the authentication information is selected, and the user who has input the authentication information corresponds to the verification information with a high similarity. Authenticated as an attached user.
[0004]
When registering collation information in the collation database, since the voice uttered by the user includes an unnecessary part such as a noise section or an unvoiced section, the section corresponding to the keyword is accurately extracted and corresponds to the keyword. It is necessary to register the feature amount of the audio signal in the section as the verification information. Since the verification information plays the role of a key in the speaker authentication apparatus, the accuracy of the section detection of the verification information greatly affects the accuracy of subsequent authentication processing.
[0005]
A method of detecting a voice section necessary for registration based on amplitude information and duration of an input signal for audio signal segmentation is disclosed (for example, “end of independent utterance by LRRabiner et al.”). An algorithm for determining the endpoint of isolated utterances ").
[0006]
On the other hand, a word spotting method that uses a similarity calculated by pattern matching such as DP matching to determine whether or not a predetermined word is present in an audio signal and the position thereof is also widely used. It has been. It is known that the word spotting method is less susceptible to noise than the method of detecting a speech section based on the amplitude information and duration of an input signal (for example, “Continuous word recognition experiment using continuous DP” by Satoru Hayami et al. And its considerations ").
[0007]
Furthermore, a method is also disclosed in which a speech recognition model representing symbols such as numerals and alphabets used as keywords based on speech input from a user is prepared and speaker recognition is performed using this speaker model ( For example, JP 2000-99090 A).
[0008]
In addition, the user's physical condition and utterance may change over time after registration of the verification information. In such a case, there is a high possibility that authentication will fail even if it is authentication information issued by the same user who registered the verification information. In this way, in order to eliminate the influence of the change in voice over time, each time authentication processing is performed, a section corresponding to the keyword is extracted from the authentication information, and the verification information is updated with the feature amount of the voice signal in that section. (For example, JP-A-57-13493).
[0009]
[Patent Document 1]
JP-A-57-13493
[Patent Document 2]
JP 2000-99090 A
[Non-Patent Document 1]
Rabina (LRRabiner et al.), "An algorithm for determining the endpoint of isolated utterances", Bell Syst. Tech. J., 1975, vol.54, pp297 -315
[Non-Patent Document 2]
Satoru Hayami, “Continuous Word Recognition Experiments and Considerations Using Continuous DP”, Transactions of the Institute of Electrical Communication, 1984, vol.J67-D, No.6, pp677-684
[0010]
[Problems to be solved by the invention]
However, in the above conventional technique, there is a problem that the verification information to be registered cannot be appropriately extracted from the audio signal acquired from the user when the verification information is registered.
[0011]
In the method of detecting the voice section based on the amplitude information and duration of the input signal, unnecessary sounds ("A ~", "E ~", etc.) generated by the user before and after the section of the verification information to be registered, and external In many cases, a section including noise is added and detected. In addition, when a consonant with a small amplitude is present at the beginning or end of the word, these consonants may be lost and detected. Furthermore, when a pause (silent section) is included in the verification information, a part of the verification information may be lost.
[0012]
Such an error in the section detection of the verification information is fatal in realizing a highly accurate authentication device. When the verification information is registered in a state where noise is mixed, the probability that the user is correctly authenticated when performing the authentication process is low. On the other hand, if a part of the verification information is registered without registration, the amount of information indicating the personal characteristics of the user is reduced, so that another person is erroneously authenticated as the user who registered the verification information. Probability increases.
[0013]
In the technique described in Japanese Patent Laid-Open No. 2000-99090, a standard speaker model is selected by a user's utterance, and a keyword speech model is created by a combination of the selected speaker models. To extract information for verification. Therefore, when the user's utterance at the time of selecting the speaker model is not clear, an appropriate speaker model is not selected, and there is a problem that a section of information for verification is erroneously detected.
[0014]
On the other hand, even when authenticating a user, if the user utters a keyword that is authentication information or if the beginning or end of the keyword is not clearly spoken, a part of the section corresponding to the keyword is missing. If a part is detected as a silent section, a section corresponding to the keyword cannot be accurately extracted from the authentication information. As a result, the accuracy of user authentication may be reduced, or authentication may be erroneous.
[0015]
In addition, when the verification information is updated at any time in order to cope with changes in the user's utterance, the verification information is updated with a keyword that is erroneously extracted from the authentication information. Adversely affected. Furthermore, the repeated update of the verification information causes a cumulative accumulation of errors, and there is a problem that the accuracy of the authentication process rapidly decreases as the number of authentication processes increases.
[0016]
The present invention has been made in view of the above-described problems of the prior art, and an object thereof is to provide a speaker authentication device and a speaker authentication program for identifying who is a speaker who can solve at least one of the above problems. .
[0017]
[Means for Solving the Problems]
  The present invention that can solve the above problemsUser uttered a keywordA database storage means for collation that stores the characteristic amount of the audio signal as collation information in association with the user, and an audio signal uttered by the user who is trying to authenticateTakeFor authenticationAudio signalAcquisition means and said authenticationAudio signalAcquired in the acquisition meansFeatures of audio signalAnd beforeNoteA user identification means for identifying a user who intends to perform the authentication by comparing application information, and a speaker authentication device comprising:The section of the section having the highest similarity between the registered voice signal acquisition means for acquiring the voice signal from the user who registers the keyword and the standard recognition model representing the keyword from the voice signal acquired by the registered voice signal acquisition means. A registered keyword section extracting means for extracting a feature quantity, and a feature quantity of the section extracted by the registered keyword section extracting means in association with a user who registers the keyword as extraction information that is not updated.Holding database storage means for extractionAnd an authentication keyword section extracting means for extracting a feature amount of a section having the highest similarity to the extraction information from the voice signal acquired by the authentication voice signal acquiring means, and a user specified by the user specifying means Database update means for updating the verification information associated with the authentication keyword section extraction means based on the feature amount extracted;It is characterized by including.
[0018]
  here,The user specifying unit specifies a user who is going to perform the authentication by comparing the feature amount extracted by the authentication keyword section extracting unit with the verification information.Is preferred.
[0019]
  Also,A keyword acquisition unit that acquires a symbol string representing a keyword, and a speech recognition model that represents each symbol acquired by the keyword acquisition unit, and a combination of these speech recognition models is used to construct the standard recognition model It is preferable to further include model construction means.
[0020]
Here, it is preferable that the keyword acquisition unit acquires a keyword using a keyboard, a pointing device, and a touch panel.
[0023]
  Further, by comparing the feature amount of the audio signal acquired by the authentication audio signal acquisition means with a part of the verification informationThe user specifying means further includes preliminary search means for narrowing down the matching information held in the matching database storage meansBeforeInformation for collation narrowed down by preliminary search meansNewsComparisonUsed forThe user who is going to authenticateMay be.
[0024]
  Another form of the present invention that can solve the above-mentioned problems isUser uttered a keywordA database for collation that stores the feature amount of the audio signal as information for collation in association with the user, and a feature amount of the audio signal used when extracting a section corresponding to the keyword from the audio signal acquired from the user.Not updatedIn a computer equipped with an extraction database to be stored as extraction information,A registration voice signal acquisition step of acquiring a voice signal from a user who registers the keyword and a section having the highest similarity between a standard recognition model representing the keyword from the voice signal acquired in the registration voice signal acquisition step. A registered keyword section extraction step for extracting a feature quantity; an extraction information registration step for registering the feature quantity of the section extracted in the registered keyword section extraction step in the extraction database as extraction information that is not updated;A voice signal uttered by the user attempting authenticationTakeFor authenticationAudio signalAn acquisition step;From the audio signal acquired in the authentication audio signal acquisition step,Information for extraction andofSimilaritymostHigh sectionFeatures as authentication informationExtractFor authenticationA keyword interval extraction step;The authentication information and the authentication informationA user specifying step of specifying a user who is going to perform the authentication by comparing the information for verification;A database update step of updating the verification information associated with the user identified in the user identification step based on the authentication information;It is characterized by executing processing includingspeakerIt is an authentication program.
[0025]
  Also,It is preferable that the user specifying step specifies a user who is going to perform the authentication by comparing the feature amount extracted in the authentication keyword section extracting step with the verification information. Also, a keyword acquisition step for acquiring a symbol string representing a keyword, a speech recognition model representing each symbol acquired in the keyword acquisition step, and combining the speech recognition models to construct the standard recognition model It is preferable to further comprise a recognition model construction step.
[0026]
Here, it is preferable that the keyword acquisition step acquires a keyword using a keyboard, a pointing device, and a touch panel.
[0029]
  Furthermore, in the speaker authentication program of the present invention, the computer includes:By comparing the feature quantity of the audio signal acquired in the authentication audio signal acquisition step with a part of the verification informationA preliminary search step for narrowing down the collation information held in the collation database;In addition, The user specifying stepBeforeThe verification information narrowed down by the preliminary search stepNewsComparisonUsed forThe user who is going to authenticateMay be.
[0030]
DETAILED DESCRIPTION OF THE INVENTION
<Authentication device>
An authentication apparatus according to an embodiment of the present invention will be described in detail with reference to the drawings. The authentication device in the present embodiment is a device that performs voice authentication for authenticating who the user is based on the voice uttered by the user.
[0031]
As shown in FIG. 1, the authentication device 100 according to the present embodiment basically includes a control unit 10, a storage unit 12, a keyword acquisition unit 14, an audio signal acquisition unit 16, a display unit 18, and a bus 20. The control unit 10, the storage unit 12, the keyword acquisition unit 14, the audio signal acquisition unit 16, and the display unit 18 are connected via a bus 20 so as to be able to transmit information to each other.
[0032]
The control unit 10 corresponds to a central processing unit (CPU) of a computer. The control unit 10 executes basic software (operation system) stored in the storage unit 12 to acquire information from the user using the keyword acquisition unit 14 and the audio signal acquisition unit 16, and uses the display unit 18 to acquire information from the user. Present information to. Further, by executing the authentication program stored in the storage unit 12, the user authentication process is performed based on the voice acquired from the user. The authentication process will be described in detail later.
[0033]
The storage unit 12 stores and holds basic software, an authentication program, and the like executed by the control unit 10. Further, information processed by the control unit 10 such as information acquired using the keyword acquisition unit 14 or the audio signal acquisition unit 16 or information presented to the user using the display unit 18 is temporarily or permanently stored. And hold. Further, the storage unit 12 stores and holds a verification database, an extraction database, and a preliminary search database used in the authentication process. The contents of these databases will be described in detail later. Information held in the storage unit 12 can be appropriately read out by the control unit 10.
[0034]
As the storage unit 12, a semiconductor memory can be used. When it is necessary to store a voice database for a large number of users, a large-capacity auxiliary storage device such as a hard disk, an optical disk, a magneto-optical disk, or a magnetic tape may be provided.
[0035]
The keyword acquisition unit 14 acquires symbols included in keywords used for authentication processing. The keyword acquisition unit 14 can be a keyboard, for example. The user inputs a symbol representing a keyword from the keyboard. The input symbol is sent to the control unit 10 for processing. Alternatively, the keyword acquisition unit 14 may be a method of selecting a symbol using a character input device such as a pointing device other than a keyboard or a touch panel.
[0036]
The audio signal acquisition unit 16 includes a microphone, an amplifier (amplifier), an analog / digital converter, and the like for acquiring audio uttered by the user. The user inputs voice using the voice signal acquisition unit 16. The voice uttered by the user is amplified by an amplifier through a microphone, converted into a digital signal by an analog / digital converter, and used for processing.
[0037]
The display unit 18 provides information necessary for processing to the user. The display unit 18 can be a display device, for example. Upon receiving an image display command from the control unit 10, the display unit 18 displays a screen prompting the user to input a keyword using the keyword acquisition unit 14 or a screen prompting the user to input a sound using the audio signal acquisition unit 16. Present. In addition, the acquired information and processing results are presented to the user. As the display unit 18, a liquid crystal display device of a touch panel, an audio output device including a speaker, and the like can be used.
[0038]
As described above, the authentication device 100 according to the present embodiment can be basically configured by an information processing device in which a microcomputer is incorporated. The authentication device 100 can be installed in the vicinity of various devices such as a door that performs user authentication processing, a safe door, and the like, and is used to authenticate who the user is by inputting voice from the user. Can do.
[0039]
Further, as shown in FIG. 2, network interfaces 22 and 24 may be further provided, and the entire authentication apparatus may be configured by separate clients 100a and servers 100b connected via a network.
[0040]
<Authentication method>
Next, a user authentication method in the present embodiment will be described. The authentication method according to the present embodiment acquires a voice signal including a keyword registered by the user for each user, extracts a section corresponding to the keyword from the voice signal, and stores the feature amount of the voice signal in that section in each database. Registration processing to be registered in the database, and verification processing for actually authenticating the user using those databases. Therefore, the registration process and the collation process will be described below.
[0041]
(registration process)
The registration process in the present embodiment is performed according to the flowchart shown in FIG. Note that the registration process in the present embodiment can be executed by the authentication apparatus of the present embodiment by programming each step of the flowchart shown in FIG. 3 and storing and holding it in the storage unit 12.
[0042]
In step S10, using the keyword acquisition unit 14, a symbol included in the keyword is acquired from the user. The control unit 10 causes the display unit 18 to display a screen that prompts the user to input a symbol string constituting a keyword to be registered. The user uses the keyword acquisition unit 14 to input a symbol group included in the keyword for identifying himself / herself into the authentication device. The input symbol group is stored in the storage unit 12.
[0043]
Keywords can consist of numeric strings. However, the present invention is not limited to this, and can be expanded to a combination of an arbitrary number of alphabets, kana characters, numbers, and other arbitrary symbols. Alternatively, some keyword candidates may be determined on the authentication device side, and the user may be allowed to select one of these candidates.
[0044]
Hereinafter, an example in which a user with user name A (hereinafter referred to as user A) registers a 4-digit number string “1234” as a keyword will be described. User A inputs a symbol included in the keyword by pressing numeric keys “1”, “2”, “3”, and “4” on the numeric keypad.
[0045]
In step S12, a recognition model is constructed based on the acquired symbol group. For each symbol that can be used as a keyword, a speech recognition model representing the symbol is registered in advance in the storage unit 12 as a model construction database, and speech recognition corresponding to the symbol acquired in step S10 from the model construction database. By extracting models and combining them, a standard recognition model representing keywords can be constructed. At this time, a recognition model can be constructed using an existing HMM (Hidden Markov Model). This recognition model is used to extract a section corresponding to a keyword from an audio signal acquired later from a user.
[0046]
Here, it is preferable that the speech recognition model stored in the model construction database is obtained by acquiring speech signals for each symbol from a large number of users and using those speech signals.
[0047]
When using a keyboard that can input alphabets, kana characters, numbers, and other symbols, it is necessary to construct a database for model construction so that a recognition model can be generated for keywords consisting of combinations of those symbols. .
[0048]
For example, as shown in FIG. 4, standard recognition models representing the keyword “1234” are extracted from the database for model construction where the speech recognition models corresponding to “1”, “2”, “3”, and “4” are extracted. As combined. Here, for example, the speech recognition model corresponding to “1” means that a speech signal corresponding to the symbol “1” uttered by a plurality of people is recognized as “1”, or a specific speaker utters a plurality of times. The voice signal corresponding to the symbol “1” is recognized as “1” using statistical processing.
[0049]
In step S14, the counter i is initialized. The counter i is used to count the number of times the user has input a keyword for registration. In this step, the counter i is set to 0.
[0050]
In step S <b> 16, an audio signal uttered by the user is acquired using the audio signal acquisition unit 16. The control unit 10 causes the display unit 18 to display a screen that prompts the user to utter a keyword, and sets the audio signal acquisition unit 16 in a voice input standby state. The user inputs an audio signal from the audio signal acquisition unit 16 by uttering the keyword. The input audio signal is analog / digital converted and stored in the storage unit 12.
[0051]
User A inputs an audio signal using the audio signal acquisition unit 16 by uttering “1234”. At this time, as shown in FIG. 5, the acquired audio signal is acquired together with the audio signal corresponding to the keyword “1234” uttered by the user, as well as the signal of an unnecessary interval such as noise from the outside or the silent interval Is done.
[0052]
In step S18, feature amounts for the entire digitally converted audio signal are extracted. The feature amount of the audio signal is preferably spectral envelope information. Spectral envelope information refers to the outline of the distribution of frequency components included in an audio signal at a certain moment. Spectral envelope information is obtained by performing spectrum analysis on a digitized signal sequence every predetermined frame width (for example, 32 milliseconds) and frame period (for example, 8 milliseconds), and using an existing LPC (Linear Predictive Coefficient) cepstrum. It can be obtained by calculation.
[0053]
However, the feature amount extracted from the audio signal is not limited to the spectrum envelope information, and may be information indicating the feature of the audio signal. For example, feature quantities such as a temporal change in the amplitude of the audio signal, an appearance period of a voiced section or an unvoiced section may be used. In the following processing, matching and extraction of audio signals are performed using the feature amount selected here.
[0054]
In the example shown in FIG. 6, the audio signal acquired from the user A is divided into a plurality of frames having a predetermined frame width every predetermined frame period, and a spectrum analysis is performed for each frame to obtain a coefficient of the 12th-order LPC cepstrum. Has been extracted.
[0055]
In step S20, it is determined whether or not the audio signal to be processed is acquired for the first time. That is, when the counter i is 0, the process proceeds to step S22, and when the counter i is other than 0, the process proceeds to step S26.
[0056]
In step S22, a section corresponding to the keyword is detected and extracted from the acquired speech signal using the recognition model constructed in step S12. Here, a word spotting method using a recognition model can be used. Using the HMM method or the like, the feature amount of the entire speech signal is compared with the feature amount of the recognition model, and a section having the highest similarity to the recognition model is extracted from the entire speech signal as a section corresponding to the keyword.
[0057]
The LPC cepstrum of the recognition model for the keyword “1234” is obtained, and as shown in FIG. 7, word spotting is executed with the LPC cepstrum of the speech signal obtained in step S18. That is, a scan for investigating the similarity to the recognition model is performed from the beginning to the end of the audio signal acquired from the user. At this time, using the HMM method, the similarity between the LPC cepstrum and the recognition model in each section of the speech signal is obtained, and the section with the highest similarity is extracted as the section corresponding to the keyword “1234”. As a result, unnecessary portions unrelated to the keyword “1234” are excluded from the audio signal acquired from the user A.
[0058]
In step S24, the feature amount of the speech signal in the section extracted as the section corresponding to the keyword is registered in the collation database, the extraction database, and the preliminary search database. The feature amount of the extracted audio signal in the section is associated with an identifier (for example, a user name) indicating the registered user, and the matching database and the extraction database are used as matching information and extraction information, respectively. Registered separately.
[0059]
In the preliminary search database, a part of the feature amount of the extracted voice signal in the section is registered as preliminary search information. That is, the preliminary search information registered in the preliminary search database is configured to have a smaller amount of information than the verification information registered in the verification database and the extraction information registered in the extraction database. For example, when the feature data of the speech signal in the section from which the verification information is extracted is all data, the preliminary search information is only partial data of the feature value of the speech signal in the extracted section. In addition, when the matching information is a multi-template, that is, a combination of feature values of speech signals in a section corresponding to a keyword extracted from a speech signal acquired a plurality of times, the preliminary search information is It may be a feature amount of one audio signal.
[0060]
When the LPC cepstrum is selected as the feature quantity, the verification information and the extraction information can be the LPC cepstrum of the audio signal in the section corresponding to the keyword. In this case, the preliminary search information registered in the preliminary search database can be obtained by thinning out the LPC cepstrum of the section corresponding to the keyword in terms of time or order.
[0061]
For example, as shown in FIGS. 8A and 8B, in the collation database and the extraction database, the LPC cepstrum in the section corresponding to the keyword “1234” in association with the user name A is the collation information and the extraction information. As registered respectively. Further, as shown in FIG. 8C, only the 0th to 8th order coefficients of the LPC cepstrum in the section corresponding to the keyword “1234” are registered in the preliminary search database in association with the user name A.
[0062]
The verification information registered in the verification database is used to authenticate the user from the voice. The extraction information registered in the extraction database is used to identify a section corresponding to the keyword in the voice signal and extract a section corresponding to the keyword from the voice signal. The preliminary search information registered in the preliminary search database is used to perform preliminary narrowing prior to user authentication using the verification information and extraction of a section corresponding to a keyword using the extraction information. .
[0063]
In addition, whether or not the registration has been correctly performed is an important factor that affects the accuracy of the subsequent collation processing. Therefore, the registered collation information, extraction information, and preliminary search information are sent to the user or the administrator. It is also preferable to perform a process for confirmation.
[0064]
In step S34, the value of the counter i is incremented by one. In step S36, it is determined whether the counter i is equal to or greater than the number of repetitions M. If the counter i is smaller than the number of times M, the process returns to step S16, and the audio signal is acquired again from the user. If the counter i is equal to or greater than the number of times M, the registration process is terminated. The number of repetitions M indicates the number of times the keyword registration process is repeated by the same user, and can be determined based on the accuracy required for the authentication process, the processing speed of the authentication apparatus, and the like.
[0065]
If it is determined in step S20 that the counter i is not 0 and the process proceeds to step S26, a section corresponding to the keyword is extracted from the audio signal based on the extraction information already registered in the extraction database. . For extraction, word spotting based on extraction information associated with a registered user can be used. A section having the highest similarity to the extraction information is extracted from the audio signal as a section corresponding to a new keyword.
[0066]
Extraction information for the user A already registered in the extraction database is selected, and word spotting is performed on the audio signal acquired from the user using the selected extraction information, which corresponds to the keyword “1234”. The section to be extracted is extracted.
[0067]
In step S28, the collation information already registered in the collation database is compared with the voice signal extracted as the section corresponding to the keyword. For comparison, existing DP matching or the like can be used. The comparison result between the two is calculated as the similarity based on the distance value between the information. In the following, the reciprocal of the distance value is calculated as the similarity. Therefore, the higher the similarity between the feature value of the extracted speech signal in the section and the verification information for the user specified by the counter i, the larger the similarity value. The calculated similarity is stored in the storage unit 12 in association with the user.
[0068]
Similarity between the LPC cepstrum of the section extracted as the section corresponding to the keyword “1234” and the LPC cepstrum as the selected collation information is selected for the collation information already registered in the collation database. The degree is calculated.
[0069]
In step S30, the calculated similarity is compared with a predetermined threshold value. If the similarity is greater than or equal to the threshold, the process proceeds to step S32. If the similarity is less than the threshold, the process returns to step S16 to acquire the audio signal again from the user.
[0070]
In step S32, the registration contents of the collation database and the preliminary search database are updated with the feature amount of the voice signal in the section corresponding to the keyword extracted in step S26. That is, the collation information already associated with the collation database associated with the user performing the registration process is updated with the feature amount of the audio signal in the section newly extracted in step S26. In addition, the preliminary search information that is associated with the user who is performing the registration process and is already registered in the preliminary search database is replaced with a part of the feature amount of the audio signal in the newly extracted section.
[0071]
For example, as shown in FIG. 9, the registration contents of the verification database and the preliminary search database are updated. Here, the collation information and the preliminary search information that are already registered in association with the user A are the LPC cepstrum of the section corresponding to the keyword newly extracted in step S26 and the 0th to 8th order coefficients of the LPC cepstrum. Replaced with each value.
[0072]
Further, if the collation information is composed of multi-templates, the template having the lowest similarity among the templates already registered may be replaced with the feature amount of the audio signal in the newly extracted section. In this case, it is preferable that the preliminary search information is a template having the highest similarity among the multi-templates of the verification information.
[0073]
It is also preferable to update with the average value of the collation information and preliminary search information already associated with the user and the feature value of the voice signal newly extracted in step S26.
[0074]
As described above, by acquiring the voice signal including the keyword from the same user a plurality of times and registering the verification information and the preliminary search information based on the plurality of voice signals, the accuracy of the user authentication process can be further improved. it can.
[0075]
As described above, according to the present embodiment, a symbol group included in a keyword to be registered by a user is acquired using an input device such as a keyboard, and a recognition model is constructed based on the symbol group. At the time of the first registration, by extracting the section corresponding to the keyword using the recognition model constructed without relying on the user's utterance as described above, unnecessary sounds and noises are generated before and after the section of the matching information to be registered. It can be prevented that the included section is added and detected, or when a consonant with a small amplitude is present at the beginning or end of the word, these consonants are missing and detected. Further, even when a pause (silent section) is included in the verification information, it is possible to prevent a part of the verification information from being lost.
[0076]
That is, the section corresponding to the keyword can be appropriately cut out from the voice signal acquired from the user, and the matching information, extraction information, and preliminary search information can be registered appropriately. As a result, the accuracy of user authentication can be improved in the following verification process.
[0077]
(Verification process)
Next, collation processing in the present embodiment will be described. The collation process is performed according to the flowchart shown in FIG. By collating each process of the flowchart shown in FIG. 10 and storing and holding it in the storage unit 12, the verification process can be realized by the authentication device.
[0078]
In step S40, a keyword is acquired as an audio signal from a user who intends to authenticate himself. The control unit 10 causes the display unit 18 to display a screen that prompts the user to utter a keyword in order to authenticate the user, and sets the audio signal acquisition unit 16 to a voice input standby state. A user who wants to perform authentication utters a keyword for identifying himself / herself in the registration process and inputs it from the voice signal acquisition unit 16. The input audio signal is analog / digital converted and stored in the storage unit 12 as authentication information.
[0079]
When the user A tries to authenticate himself / herself, the user A utters his / her keyword “1234” toward the microphone and inputs a voice signal as authentication information to the authentication apparatus. At this time, the acquired authentication information includes the keyword “1234” uttered by the user A, as well as an unnecessary sound, an external noise, and an audio signal of an unnecessary section such as a silent section.
[0080]
In step S42, a feature amount is extracted from the digitally converted authentication information. The feature amount of the authentication information is the same type as the feature amount registered in the collation database, the extraction database, and the preliminary search database in the registration process. For example, when the coefficient value of the LPC cepstrum is registered in each database, the LPC cepstrum is obtained as a feature amount from the authentication information. The details of the process here are the same as in step S18, and thus the description thereof is omitted.
[0081]
Here, the authentication information acquired from the user A is divided into a plurality of frames having a predetermined frame width every predetermined frame period, and a spectrum analysis is performed for each frame to extract a 12th-order LPC cepstra coefficient. The
[0082]
In step S44, a preliminary search for the authentication information is performed using the preliminary search information registered in the preliminary search database. Using the existing word spotting method or the like, the preliminary search information included in the preliminary search database is sequentially compared with the feature amount of the authentication information acquired from the user. Only a predetermined number of users C are preliminarily selected in ascending order of similarity calculated using the DP matching method or the like.
[0083]
For example, when the LPC cepstrum coefficients 0 to 8 of the keyword for each user are registered as preliminary search information in the preliminary search database, the 0 to 8th order coefficients of the LPC cepstrum of the authentication information obtained in step S42. Are matched with each preliminary search information included in the preliminary search database, and the preliminary search information in the authentication information is associated with the preliminary search information in descending order of similarity. A predetermined number of user names are selected. Identification numbers are assigned to the selected users in order from 1.
[0084]
Assuming that the number of people C to be extracted in the preliminary search is 3, three preliminary search information is selected in descending order of similarity calculated using the DP matching method or the like, and is associated with the selected preliminary search information. Users are preliminarily selected. In the following description, it is assumed that users A, B, and C are selected, and identification numbers 1, 2, and 3 are assigned to the users A, B, and C, respectively.
[0085]
In this way, by performing narrowing down of users by using preliminary search information having a smaller amount of information than collation information and extraction information, it is possible to reduce the burden of subsequent processing.
[0086]
In step S46, the counter j is initialized. The counter j is used to count the number of users that have been verified, and the counter j is set to 1 in this step.
[0087]
In step S48, the section corresponding to the keyword is extracted from the authentication information using the extraction information registered in the extraction database. Using a word spotting method or the like, a section having the highest similarity with the extraction information associated with the user that is preliminarily selected in step S44 and specified by the value of the counter j is extracted from the authentication information. The extracted feature quantity of the audio signal in the section is stored in the storage unit 12 in association with the user specified by the counter j.
[0088]
If the users A, B, and C are narrowed down in step S44 and the identification numbers 1, 2, and 3 are assigned to them, and the counter j is 1, the LPC cepstrum corresponding to the user A from the extraction database. Is selected, and a section having the highest similarity with the coefficient value of the LPC cepstrum is extracted from the authentication information. If the counter j is 2, extraction is performed using the coefficient value of the LPC cepstrum corresponding to the user B, and if the counter j is 3, extraction is performed using the coefficient value of the LPC cepstrum corresponding to the user C.
[0089]
In step S50, the similarity between the speech signal in the section extracted in step S44 and the collation information registered in the collation database is calculated. Using the DP matching method or the like, the feature amount of the extracted voice signal in the section is compared with the verification information for the user specified by the counter j, and the similarity between the information is calculated. The calculated similarity is stored in the storage unit 12 in association with the user.
[0090]
When the users A, B, and C are narrowed down in step S44 and the identification numbers 1, 2, and 3 are assigned to them, if the counter j is 1, the LPC cepstrum corresponding to the user A from the collation database. Coefficient values are selected, and the similarity to the coefficient value of the LPC cepstrum in the section extracted in step S44 is obtained. If the counter j is 2, the degree of similarity is obtained using the coefficient value of the LPC cepstrum corresponding to the user B, and if the counter j is 3, the coefficient value of the LPC cepstrum corresponding to the user C is obtained.
[0091]
In step S52, the value of the counter j is incremented by 1. In step S54, it is determined whether or not the value of the counter j is equal to or greater than the number C of users detected in advance. If the counter j is smaller than the number of users C, the process returns to step S48, and the process is repeated for the user to which the next counter j value is assigned. If the counter j is equal to or greater than the number of users C, the process proceeds to step S56.
[0092]
Through the processing so far, the similarity between the authentication information and the verification information of each user is obtained for each of the users A, B, and C.
[0093]
In step S56, the similarity stored in the storage unit 12 is read, and the C similarity having the highest value, that is, the highest similarity is selected. The value is compared with a preset threshold. If the similarity is greater than the threshold, the process proceeds to step S58. If the similarity is equal to or less than the threshold, the process proceeds to step S60.
[0094]
In step S58, the user who is performing the authentication process is authenticated as the user corresponding to the similarity having the largest value. Along with the authentication, a screen indicating that the user has been authenticated may be displayed on the display unit 18 or a door lock may be unlocked.
[0095]
When the user A is going to authenticate, the similarity between the user A, B, and C and the user A's verification information is the highest, and the similarity to the user A exceeds the threshold. On the other hand, when a user who has not performed keyword registration processing performs authentication, the similarity to all the preselected users does not exceed the threshold value, and the user is not authenticated.
[0096]
In step S60, processing is performed when the user is not authenticated. For example, you may perform the process which displays the screen which shows that a user was not authenticated on the display part 18. FIG.
[0097]
In step S62, as the user is authenticated, the verification information and the preliminary search information regarding the authenticated user are updated. That is, the registration contents of the collation database and the preliminary search database are updated with the feature amount of the voice signal in the section corresponding to the keyword held in the storage unit 12 in association with the user who is authenticated in step S48.
[0098]
For example, the verification information that is associated with the authenticated user and already registered in the verification database is replaced with the LPC cepstrum of the audio signal that is stored in association with the authenticated user in step S58. If the collation information is composed of multi-templates, the template having the smallest similarity among all templates and the LPC cepstrum of the extracted audio signal are replaced.
[0099]
Further, the preliminary search information that is associated with the authenticated user and already registered in the preliminary search database is replaced based on the LPC cepstrum of the audio signal held in association with the authenticated user. For example, when the matching information is all coefficient values of the LPC cepstrum, the preliminary search information is replaced with a part of coefficient values of the LPC cepstrum. In addition, if the collation information is composed of a multi-template of LPC cepstrum coefficient values, it may be replaced with a template having the highest similarity among these templates.
[0100]
Also, the average of the collation information and preliminary search information already associated with the authenticated user and the LPC cepstrum of the audio signal associated with the authenticated user and some coefficient values of the LPC cepstrum It is also preferable to update by value.
[0101]
In this way, by updating the authenticated user verification information and preliminary search information, it is possible to suppress a decrease in the accuracy of authentication with respect to changes over time such as the physical condition of the user and the manner of utterance.
[0102]
In addition, since the extraction information registered in the extraction database is not updated, the process of extracting the section corresponding to the keyword from the authentication information is performed invariably from the time of keyword registration. Therefore, it is possible to reduce the influence of cumulative accumulation of errors in verification information and preliminary search information due to repeated authentication processing. In other words, it is possible to change the utterance of the user over time by separately performing the segmentation of the keyword section using the extraction information that is not updated and the final authentication of the user using the verification information updated for each authentication. Corresponding high-accuracy authentication processing can be realized, and accumulation of errors due to the update of verification information can be reduced.
[0103]
In the present embodiment, preliminary narrowing down of users is performed using preliminary search information, but the preliminary search processing may be omitted.
[0104]
<Modification 1>
Next, a modification of the embodiment of the present invention will be described. This modification can be executed using the authentication device in the above embodiment. This modification is also roughly divided into a registration process and a collation process. Since the registration process is the same as the above process, only the collation process will be described below.
[0105]
The matching process in the modification is performed according to the flowchart shown in FIG. By collating each process of the flowchart shown in FIG. 11 and storing and holding it in the storage unit 12, the verification process can be realized by the authentication device. Here, the same reference numerals are assigned to the steps of performing the same processing as the verification processing in the above embodiment, and the description thereof is omitted.
[0106]
In step S64, the feature amount of the audio signal extracted in step S42 is compared with the verification information registered in the verification database. Using a word spotting method or the like, a section having the highest similarity with the verification information associated with the user selected in step S44 and specified by the counter j is extracted from the authentication information. Further, using the DP matching method or the like, the feature amount of the extracted speech signal in the section is compared with the verification information for the user specified by the counter j, and the similarity between the information is calculated. Therefore, the higher the similarity between the feature value of the extracted speech signal in the section and the verification information for the user specified by the counter j, the greater the similarity value. The calculated similarity is stored in the storage unit 12 in association with the user.
[0107]
For example, when the users A, B, and C are narrowed down in step S44 and the identification numbers 1, 2, and 3 are assigned to the users A, if the counter j is 1, it corresponds to the user A from the collation database. The verification information is selected, the section having the highest similarity with the verification information is extracted from the authentication information, and the similarity between the feature amount of the section and the verification information is calculated. If the counter j is 2, the similarity is calculated using the collation information corresponding to the user B, and if the counter j is 3, the collation information corresponding to the user C is used.
[0108]
That is, in this modification, the degree of similarity between the authentication information and the verification information for each user is calculated without using the extraction information registered in the extraction database. And based on those similarities, a user authentication process is performed in steps S56 to S60.
[0109]
In step S66, the section corresponding to the keyword is extracted from the audio signal using the extraction information associated with the user corresponding to the maximum similarity, that is, the authenticated user. Extraction information associated with the authenticated user is selected from the extraction database, and the feature amount of the voice signal extracted in step S42 is compared with the extraction information using a word spotting method or the like, and the extraction information is extracted. The section having the highest similarity with the information is cut out as a section corresponding to the keyword.
[0110]
In step S62, the verification information and preliminary search information associated with the authenticated user are updated with the feature amount of the section corresponding to the keyword.
[0111]
According to this modification, since the section corresponding to the keyword is extracted only for the authenticated user, the processing load of the user authentication processing in steps S46 to S54 can be reduced. As a result, it is possible to shorten the waiting time until the authentication result is obtained after the user inputs voice.
[0112]
【The invention's effect】
ADVANTAGE OF THE INVENTION According to this invention, the keyword used for authentication can be registered correctly, and the speaker authentication which is hard to be influenced by the time-dependent change of a user's utterance is realizable.
[Brief description of the drawings]
FIG. 1 is a block diagram showing a configuration of an authentication apparatus according to an embodiment of the present invention.
FIG. 2 is a block diagram showing another configuration of the authentication device according to the embodiment of the present invention.
FIG. 3 is a diagram showing a flowchart of speaker authentication registration processing in the embodiment of the present invention.
FIG. 4 is a diagram illustrating an example of construction of a recognition model.
FIG. 5 is a diagram illustrating an example of an audio signal acquired from a user.
FIG. 6 is a diagram for explaining an example of feature amount extraction from an audio signal;
FIG. 7 is a diagram illustrating an example in which a word spotting method is applied to an audio signal acquired from a user.
FIG. 8 is a diagram illustrating an example of registered contents of a collation database, an extraction database, and a preliminary search database.
FIG. 9 is a diagram for explaining an example of updating a collation database and a preliminary search database;
FIG. 10 is a diagram showing a flowchart of verification processing for speaker authentication in the embodiment of the present invention.
FIG. 11 is a diagram illustrating a flowchart of verification processing for speaker authentication according to a modification of the embodiment of the present invention.
[Explanation of symbols]
DESCRIPTION OF SYMBOLS 10 Control part, 12 Memory | storage part, 14 Keyword acquisition part, 16 Audio | voice signal acquisition part, 18 Display part, 20 Bus, 22, 24 Network interface, 100 Authentication apparatus, 100a client, 100b server.

Claims (5)

Collating database storage means for a user to hold in association with the user as the verification information feature amount of the voice signal uttered keywords, authentication audio signal acquiring means acquire the speech signal by the user utters that tries to authenticate If, speaker authentication and a user identification means for identifying a user who attempts the authentication by comparing the feature quantity of the acquired audio signal and before KiTeru if information in the authentication voice signal acquisition means A device,
Registered voice signal acquisition means for acquiring a voice signal from a user who registers the keyword;
A registered keyword section extracting unit that extracts a feature quantity of a section having the highest similarity to a standard recognition model representing the keyword from the voice signal acquired by the registered voice signal acquiring unit;
Extraction database storage means for retaining the feature quantity of the section extracted by the registered keyword section extraction means in association with a user who registers the keyword as extraction information that is not updated ;
An authentication keyword section extraction unit that extracts a feature amount of a section having the highest similarity to the extraction information from the voice signal acquired by the authentication voice signal acquisition unit;
Database updating means for updating matching information associated with the user specified by the user specifying means based on the feature quantity extracted by the authentication keyword section extracting means;
A speaker authentication device comprising: The speaker authentication device according to claim 1,
The speaker authentication apparatus characterized in that the user specifying means specifies a user who is going to perform the authentication by comparing the feature amount extracted by the authentication keyword section extracting means with the verification information. In the speaker authentication device according to claim 1 or 2,
Keyword acquisition means for acquiring a symbol string representing a keyword;
A recognition model construction means for obtaining a speech recognition model representing each symbol obtained by the keyword obtaining means, and combining the speech recognition models to construct the standard recognition model;
A speaker authentication device , further comprising: The speaker authentication device according to any one of claims 1 to 3 ,
Preliminary search for narrowing down the collation information held in the collation database storage unit by comparing the feature amount of the voice signal acquired by the authentication voice signal acquisition unit with a part of the collation information Further comprising means,
The user identification means, before Symbol speaker authentication apparatus characterized by specifying the user to be performed the authentication by using the comparison collation information narrowed down by the preliminary search means. A database for collation that holds the feature quantity of the audio signal from which the user uttered the keyword as a collation information in association with the user, and an audio signal used when extracting a section corresponding to the keyword from the audio signal acquired from the user In a computer comprising an extraction database that retains the feature quantity of the information as extraction information that is not updated ,
A registered audio signal acquisition step of acquiring an audio signal from a user who registers the keyword;
A registered keyword section extraction step for extracting a feature amount of a section having the highest similarity to a standard recognition model representing the keyword from the voice signal acquired in the registered voice signal acquisition step;
An extraction information registration step of registering the feature quantity of the section extracted in the registered keyword section extraction step in the extraction database as extraction information that is not updated;
And authentication audio signal acquiring step get the audio signal that the user utterance attempting to authenticate,
Authentication keyword block extraction step of similarity between the extracted information from the speech signal to extract a feature value of the highest section as authentication information acquired in the authentication voice signal acquisition step,
A user specifying step of specifying a user who is going to perform the authentication by comparing the authentication information and the verification information;
A database update step of updating the verification information associated with the user identified in the user identification step based on the authentication information;
A speaker authentication program for executing a process including:

Images